The following is new to Python 3.9, and I consider the implementation incomplete. I have code that works for Python 3.8 and before, but not for Python 3.9:

"Class methods can now wrap other :term:`descriptors <descriptor>` such as :func:`property`."

#8405
https://bugs.python.org/issue19072

As implemented, classmethod does not correctly wrap descriptors that mimic classmethod. Previously, __get__of wrapped objects wasn't invoked by classmethod, so it was safe to have an object with both call and get behave like a classmethod. Now, classmethod calling get first gives incorrect results.

Here is a minimal example:

from types import MethodType


class myclassmethod:
    def __init__(self, func):
        self.func = func

    def __call__(self, cls):
        return self.func(cls)

    def __get__(self, instance, owner=None):
        if owner is None:
            owner = type(instance)
        return MethodType(self, owner)


class A:
    @myclassmethod
    def f1(cls):
        return cls

    @classmethod
    @myclassmethod
    def f2(cls):
        return cls


assert A.f1() is A
assert A.f2() is A  # <-- fails in 3.9, works in 3.8 and before

This pattern would typically be used to do something extra in __call__.

For the sake of discussion, let's call the two arguments to __get__ "instance" and "owner". Typically, "instance" is an instance of "owner", or, equivalently, "owner" is the type of "instance". If "owner" is None, it is generally assumed to be the type of "instance".

In bpo-19072 (and gh8405), classmethod was changed to call obj.__get__(owner) if the wrapped object "obj" has get. Notice that only the "instance" argument is provided. Moreover, the type owner is passed as the "instance" argument. This means that the "owner" argument (which is None) will be assumed to be the type of the "instance" argument, which is the type of the owner type. This is wrong. The "owner" argument should be owner.

I believe it would be better for classmethod to call obj.__get__(owner, owner) if "obj" has get.

This is kind of difficult to explain. I will make a PR with more informative tests shortly. Here is the simple diff to make the above example pass:

diff --git a/Objects/funcobject.c b/Objects/funcobject.c
index bd24f67b97..74f9167566 100644
--- a/Objects/funcobject.c
+++ b/Objects/funcobject.c
@@ -739,7 +739,7 @@ cm_descr_get(PyObject *self, PyObject *obj, PyObject *type)
         type = (PyObject *)(Py_TYPE(obj));
     if (Py_TYPE(cm->cm_callable)->tp_descr_get != NULL) {
         return Py_TYPE(cm->cm_callable)->tp_descr_get(cm->cm_callable, type,
-                                                      NULL);
+                                                      type);
     }
     return PyMethod_New(cm->cm_callable, type);
 }

Since I consider the new behavior to have introduced a regression, I think this change should be applied to both 3.9 and 3.10.

Cheers!

Thanks! ✨ 🍰 ✨

New changeset b83861f by Łukasz Langa in branch 'main':
bpo-42073: allow classmethod to wrap other classmethod-like descriptors (bpo-27115)
b83861f

New changeset 2ce8af3 by Miss Islington (bot) in branch '3.10':
bpo-42073: allow classmethod to wrap other classmethod-like descriptors (GH-27115) (GH-27162)
2ce8af3